Phonetic minimization of the text corpus in Belarusian for the speech synthesis system training

Abstract

The most modern speech synthesis systems are based on the corpus-based method. The corpus-based method, unlike previously popular compilation method, uses natural speech database that does not consist of separate specially selected elements of compilation, but represents the corpus of phonograms of natural speech. Large amounts of text and corresponding audio information, which represents a significant challenge for so-called under-resourced languages, which include Belarusian, are required to achieve high-quality synthesized speech in this approach. In this case, a common approach is to use phonetic minimization, special selection of texts, when the amount of text corpus is maximally reduced, but at the same time phonetic fullness is preserved. The article discusses the information about the nature and the functioning the corpus-based method of sound signal generation in speech synthesis systems, provides a detailed overview of the approaches to the formation of text and speech corpuses, required for speech generation by the corpus-based method. The second half of the work is devoted to the description of the elaborated algorithm of the text corpus phonetic minimization in Belarusian language, as well as technical and linguistic resources used to implement it. A description of the developed software prototype as well as a description of the series of experiments on phonetic minimization are given to demonstrate the efficiency of the algorithm.